Flervalsinlärning i talseparering. Från många till varierande talares separation

This report deals with the speech processing source separation task and introduces Multiple Choice Learning (MCL) to tackle the assignation ambiguity between predictions and ground-truth-separated signals. This inherently ambiguous task is customarily solved using Permutation Invariant Training (PIT) which finds the best assignation at the cost of a computationally expensive procedure. I demonstrate experimentally that MCL, originally introduced to tackle ambiguous tasks, matches the performances of PIT while being computationally advantageous in many-speaker settings. The variable number of speakers separation task is another approach tackled by introducing a scoring model that helps to quantify how "relevant" a prediction is during the separation. It is shown that the scoring model allows improvements in speech separation for 2 and 3-speakers based on a 5-speaker pre-trained model. Two new metrics W_EMD and T_EMD using scores are introduced and enable a better understanding of what the different predictions contain in a variable number of speaker settings. Finally, one can observe separation improvements while using the scoring, which is not detached from the separation.